Information Technology Reference
In-Depth Information
architecture is optimal, and,
just as important,
it
is far from exhausting its
possibilities.
The situation changes radically when it comes to the development of neural
information-logical devices.
The basic principles of the neural network paradigm are fundamentally different
from the principles of the von Neumann paradigm. At the core of the von Neumann
paradigm is the notion of a program with unalterable structure. Changing the
program, even a single operator, leads to its disintegration and renders it unable
to perform its functions.
By contrast, the ideology of neural networks, even the initial approach of
M
Culloch and Pitts, is fundamentally different in that it provides for the possibility
of small changes in the structure of the network. The concept of variable weights of
neurons allows to vary them in a certain range of values, without causing qualitative
changes in the operating mode of the network.
It is precisely this feature where the discrete semiconductor implementation of
neural networks fundamentally contradicts to their nature.
Removing or adding even a single transistor to a planar circuit, in general, leads
to loss of its function (information redundancy is not considered here as it does not
change the overall conclusions). Therefore, gradual adaptation of the circuit toward
a more efficient solution of a specific problem encounters major difficulties when
using semiconductor planar technology.
At the same time, known biological systems with neural network architecture are
built of basic molecular fragments that are qualitatively different from semicon-
ductor components (transistors). One of the main and probably most important
features of such systems is the structural redundancy of molecular objects with
respect to their functions. Thus, biopolymer molecules of protein enzymes play an
important role in the functioning of biological systems. Their structure is a combi-
nation of functional groups, defining the function of the enzyme, and an extended
(polypeptide) “tail.” A remarkable feature of this structure is that the removal of
even relatively large fragments from the tail leads to only marginal change of
enzyme function.
Functional redundancy of biological systems also manifests itself when the
change of dynamic characteristics of the system in a fairly wide range does not
lead to qualitative changes of dynamics, i.e., to transition to another regime. This
can be defined as a dynamic redundancy of the system (see below).
In general, structural and (or) dynamic redundancy of elements (molecular
fragments) used to construct logic information systems constitutes the basis of
their variability. This, in turn, should be the basis of evolutionary selection.
Therefore, such source elements may be, in principle, utilized to build
information-logical devices, capable of gradually learning the most efficient solu-
tion of the problem in the course of the decision process.
с
